Exploratory Data Story – Soil Moisture LSTM
soil moisture automation flows from sensor nodes into our lorawan database…
- placeholders now: swap in field hardware and pipeline schematic when you have them
start with the raw ingest to show the season span and anomalies we cleaned…
- eleven week window per plot from july nineteen to october thirteen twenty twenty three
- raw versus cleaned snippets spotlight duplicates and flatlines we resolved
calm the probe noise while keeping irrigation pulses intact for agronomy decisions…
- pchip interpolation plus savitzky–golay smoothing keeps peaks without lagging response
note the savitzky golay smoothing expression so collaborators can rerun it…
record the drying rate transform because the model needs velocity not just level…
layer irrigation signals the same way agronomists track water budget…
- raw totals, log scale, rolling sum, and actuator flag sit side by side
write down the irrigation transforms that drive our features for reproducibility…
reveal the feature buffet and the documented cuts so trust stays high…
- kept buckets cover soil layers, temporal encoding, weather load, irrigation context
display the seven day window feeding the ninety six hour forecast before results…
- diagram makes the sliding window and horizon obvious in seconds
plot two zero zero three shows surface tracking yet deeper layers lag after irrigation…
- discuss irrigation sensitivity still missing at depth
plot two zero one four stays biased high because dry down exemplars were missing in training…
- explain the imbalance across plots here
plot two zero one five lags forty eight hours and exposes the fold reuse bug plus irrigation imbalance…
- highlight how repairing the split will sharpen responsiveness
stress test injection reveals frozen features and off manifold inputs we must fix…
- recompute engineered features before injecting and stay within scaler range next time